It is becoming increasingly clear that quantum computers will not have a single moment in which they are clearly superior to classic hardware. Instead, we will likely see them become useful for a limited number of problems, and then gradually expand to cover an increasing range of computations. The question becomes obvious, where is the benefit seen first?
The quantum computing startup Rigetti now has a whitepaper that at least theoretically shows a case in which quantum hardware should offer an advantage. And it’s actually useful: it replaces a neural network that is used to analyze weather data.
What’s the weather like?
The problem the folks at Rigetti have been dealing with is taking a subset of the weather data and figuring out what the rest of it looks like. Many areas of the planet do not have good coverage and we only get partial information about local conditions. And when we have things like airliners traversing these remote areas, we often want a more complete picture of the conditions there.
To cope with this, people have trained neural networks in regions where we have more complete weather data. After training, the system could be fed partial data and derived from it what the rest would likely be. For example, the trained system can create a probable weather radar map that uses things like satellite cloud images and data about lightning strikes.
This is exactly what neural networks are good at: Recognizing patterns and drawing conclusions about correlations.
What caught the Rigetti team’s attention is the fact that neural networks can also be mapped well on quantum processors. In a typical neural network, one layer of “neurons” perform operations before passing their results on to the next layer. The network “learns” by changing the strength of the connections between units in different layers. On a quantum processor, each qubit can perform the equivalent of an operation. The qubits also share connections with each other, and the strength of the connection can be adjusted. So it is possible to implement and train a neural network on a quantum processor.
Could be better
Conveniently, some researchers at Google have worked out a metric that enables the comparison of AIs implemented on classical and quantum hardware. And Rigetti built a 32-qubit quantum processor so that he could do the comparison. And based on this metric, there are at least some cases where a quantum system should outperform a classic one.
However, it remains unclear exactly which cases are involved. Therefore, the researchers experimented with a number of ways to use their quantum processor as part of a mixed quantum / classical system. They found that the system was more or less successful on various aspects of the weather data. For example, when they used the quantum processor to reconstruct lightning data, they found that it worked better at lower altitudes but was generally comparable to the classical neural network.
In a separate test, they simply replaced the neural network with qubits. For lightning data, the quantum version outperformed the classic. The tables were turned when it was tested with satellite data, where classic systems were more accurate.
It is important to emphasize that at no point did the quantum system show any real performance advantage over existing methods of performing this type of weather analysis; The important takeaway here is the indication that better performance is possible. As the Rigetti researchers themselves state: “These results are the first evidence that data is in the real world [machine-learning] Problems – here high-dimensional weather data – can have a structure that is theoretically compatible with quantum advantages. “
Their ability to perform parts of the analysis on quantum hardware with decent results shows that there is no obstacle to integrating quantum methods into this type of analysis as well. While this isn’t the kind of breakthrough that gets attention, it is the kind of hard work that is required for quantum computers to reach their potential.